A schema draft helps route questions across private data owners
In plain words
When several people or teams keep their own tables private, an AI coordinator may need information from more than one owner to answer a question. Asking everyone separately can fail because no one gets the matching details returned by someone else.
The paper introduces OAR, which reads public table descriptions and a draft query to choose between asking owners independently or one at a time. Across ten tested settings, OAR was not significantly less accurate than the best fixed approach in any setting.
The research
When table rows stay with their owners, a coordinator may need to combine results without being allowed to see the underlying tables.
It remains unclear whether it should ask owners independently or sequentially, especially when one owner’s returned rows may contain keys needed to find matching rows elsewhere.
OAR routes each question to parallel or sequential querying from a SQL draft over public schemas, before owners reveal private rows.
The authors compare fan-out, where owners answer independently, with relay, where a planner can use earlier returned rows. Exact accuracy measures the share of questions answered with the reference rows, counting abstention as correct only for unanswerable questions.
Fan-out usually leads when one owner holds the needed tables, but answers at most 4.3% correctly when two owners must be joined; relay reaches 11.6–47.9%. This reversal supports choosing a method based on ownership.
The evidence comes from English questions over relational tables and three model families, with Relay+ tested using two. OAR also assumes a public table ownership catalog and depends on the quality of its model-written draft.
Analysis
Deep
- Original paper authors
- Zhexi Feng
- Submitted to arXiv
- License
- CC BY 4.0